Need to regress residuals on orbit positions

[Ian Culverwell]

Stig Syndergaard (DMI) discovered that the problematic difference between ifort/linux and xlf95/HPC when processing RS data with ropp_pp_occ_tool could be explained by the different accuracies in the orbit fitting routine ropp_pp_regression on the two platforms. See #301 for details. The solution is to carry out a further fit to the residuals (data-fit) of the first fit. The results are then much more similar.

Stig has written a routine ropp_pp_residual_regression (attached in ropp_pp_utils.f90), which should be called every time (=4 I think, but needs checking) that ropp_pp_regression is called to fit satellite orbit co-ordinates.

[Ian Culverwell]

ropp_pp_utils.f90_SSY

[Ian Culverwell]

ropp_pp_preprocess_grasrs.f90_SSY

[Ian Culverwell]

There are three elements to this change:

• Inclusion of ropp_pp_residual_regression in ropp_pp_utils.f90;
• Fit orbit with non-missing data only;
• Regress residual orbit positions whenever ropp_pp_regression is used to fit the orbits. This occurs in:
  • ropp_pp_satellite_velocities.f90
  • ropp_pp_preprocess_grasrs.f90
  • ropp_pp_bangle2phase.f90
  • ropp_pp_radiopotic_analysis.f90

The second change has not been carried out, as it is difficult to handle all possible cases of missing data. But #301 shows that this infelicity has little impact. The other two changes have been made at r4209.

[Ian Culverwell]

The case for this change has been exhaustively made in #301. But the impact on the problematic 3rd profile of the (old) IT-PP-05.nc dataset has been rechecked. In short, there is a significant difference betwen the results of ropp_pp_occ_tool when run with this profile on ifort/linux and xlf95/AIX:

Difference between compilers at ROPP7.1

Regressing the residual orbit positions (and the other ROPP8.0 changes which have minor effect) makes little change to the ifort results but a large one to the xlf95 results:

Difference ROPP8.0 and ROPP7.1 on both compilers

This means, finally, that both compilers/platforms have very similar behaviour on this profile:

Difference between compilers at ROPP8.0

[Ian Culverwell]

Closing ticket.